TL;DR
This paper explores the application of Vision Transformer (ViT) and its variant, Swin Transformer, for landslide susceptibility mapping (LSM), demonstrating their superior ability to capture global spatial features compared to traditional CNNs and machine learning models. The study shows that Transformer models, particularly Swin-T, achieve higher accuracy and better generalization, offering a promising new direction for more reliable landslide risk assessment.
What Problem Does This Paper Address?
Landslide susceptibility mapping (LSM) is crucial for disaster prevention and mitigation, aiming to predict where landslides are most likely to occur. Traditional methods, including machine learning (ML) and Convolutional Neural Networks (CNNs), have limitations. CNNs, while effective at feature extraction, primarily focus on local information due to their fixed convolutional kernels, potentially missing broader spatial relationships critical for understanding landslide triggers. ML models often rely heavily on pre-existing knowledge and can struggle with high-dimensional data and complex spatial patterns. This paper addresses the need for LSM models that can effectively capture both local and global spatial dependencies, leading to more accurate and generalizable predictions.
Methods
The researchers investigated four models for LSM: Support Vector Machine (SVM) as a representative ML model, ResNet50 as a standard CNN, and two Transformer-based models: Vision Transformer (ViT-B/16) and Swin Transformer (Swin-T).
- Data-Driven Approach: The study employs a data-driven approach, using historical landslide data and various conditioning factors.
- Feature Engineering: Eleven landslide conditioning factors were selected, including elevation, slope, aspect, curvature (plan and profile), rainfall, lithology, land cover, distance to river, distance to fault, and NDVI.
- Transformer Models:
- ViT: This model treats images as sequences of patches, leveraging self-attention mechanisms to capture global relationships across the entire image.
- Swin Transformer: An evolution of ViT, Swin introduces a hierarchical structure and shifted window mechanism, allowing for efficient local and global feature extraction, similar to CNNs but with enhanced global context.
- Model Training and Evaluation: The dataset was split into training, validation, and testing sets. Models were trained and evaluated using metrics like accuracy, F1-score, and AUC (Area Under the ROC Curve).
Datasets
The study utilized data from Pingwu County, Sichuan Province, China, a region prone to landslides.
- Landslide Data: 150 landslide points were collected from the Resource and Environmental Science and Data Centre of the Chinese Academy of Sciences, supplemented with satellite imagery, government reports, and Google Earth.
- Landslide Conditioning Factors:
- DEM: 10m resolution Digital Elevation Model.
- Fracture Zone Data: From the Chinese Academy of Surveying and Mapping Sciences.
- Lithology Data: From the Resource and Environment Science and Data Centre.
- Rainfall Data: Calculated average rainfall (1991-2020) with 30m resolution.
- Landcover Data: From the China Earth Big Data Science Project.
- NDVI Data: Derived from Google Earth Engine using Landsat observations.
- Data Preprocessing: All data were unified into GeoTIFF format with a 10m spatial resolution, resulting in 11 landslide conditioning factors. Input data for ResNet50, ViT-B/16, and Swin-T were formatted as 224x224x11 dimensional matrices.
Key Insights and Ideas
- Global Feature Extraction is Key: The study strongly suggests that Transformer models' ability to capture global spatial dependencies is a significant advantage over CNNs' local focus for LSM. This is crucial because landslide occurrences are influenced by a wider geographical context than just immediate surroundings.
- Swin-T's Hybrid Approach Excels: Swin Transformer, by integrating CNN-like hierarchical feature extraction with Transformer's global attention, appears to strike an optimal balance, leading to the best performance.
- Transformer Models Generalize Better: The ViT-B/16 model, in particular, showed better alignment with actual landslide events and demonstrated stronger generalization ability compared to SVM and ResNet50, even in areas where historical landslide data might be sparse.
- Feature Map Variance Matters: Analysis of feature map variance revealed that Swin-T maintains lower variance, contributing to more stable and accurate predictions. ViT also shows promising trends, while CNNs exhibit increasing variance.
Key Results
- Performance Metrics: In the testing dataset, Swin-T achieved the highest accuracy (88.33%) and AUC (97.22%), closely followed by ViT-B/16 (Accuracy: 88.33%, F1-score: 88.89%, AUC: 96.56%). ResNet50 and SVM performed comparably but lower than the Transformer models.
- Spatial Distribution: The spatial distribution of predicted landslide susceptibility from ViT-B/16 and Swin-T was found to be more consistent with actual landslide events and geological features (like fault zones) than SVM and ResNet50.
- Landslide Point Distribution: A higher proportion of landslide points were found in the high and very high susceptibility zones predicted by all models, with Swin-T and ViT-B/16 showing particularly strong agreement.
Limitations and Future Work
- Computational Cost: While ViT-B/16 offers high accuracy, it has deeper layers and more parameters, leading to slower processing speeds compared to SVM, which can limit its practical application in real-time scenarios.
- Input Data Size: The fixed input size of 224x224 for Transformer models might limit the spatial resolution of the evaluation factors, potentially leading to redundant information extraction. The optimal window size for input data remains an open question.
- Topographic Relief: The study area had relatively gentle topography. The contribution of elevation to landslide interpretation in steeper terrains needs further investigation.
- Future Directions: The authors suggest exploring the advantages of global vs. local feature extraction models further, combining feature extraction with various ML methods, and investigating the optimal input data size for Transformer models in LSM. They also plan to compare different feature extraction methods and their effectiveness.
Summary updated